Pruning and acceleration of deep neural networks
Date
2020
Authors
Thivagara Sarma, Janarthanan, author
Pouchet, Louis-Noël, advisor
Rajopadhye, Sanjay, committee member
Pasricha, Sudeep, committee member
Anderson, Chuck, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Deep neural networks are computational and memory intensive applications. Many network pruning and compression solutions has been introduced to deploy inference of large trained models in limited memory and time critical systems. We proposed a new pruning methodology that assigns significance rank to the operations in the inference program and for a given capacity and operation budget, generate only the important operations to do the inference. Our approach has shown that, in many classical feed forward classification networks we can maintain almost the same accuracy as the original inference by executing less than half of the operations in the original program. We also proposed a methodology to improve the effective implementation of the output sparse computation, controllable by a threshold variable.
Description
Rights Access
Subject
compression
pruning
acceleration
SIMD
deep neural networks